On Gabor Wavelet Feature Extraction Technology And Its Applied Research In Target Recognition

The automatic target recognition is a pop issue in the computer vision area, that has been used abundantly in many fields such as image, biology, industry and so on, especially in military circles. During ATR's taches, extracting feature of target image should be the most difficult and important one. The aim of this paper is to enhance the robustness of corresponding Gabor algorithm and to weaken the effect of negative factors.Firstly, some main classifiers are presented, and some advantage and disadvantage of BP network which belong to the ANN are analysed. An improved method whose step extent shift is given to optimize the BP net, which supplies a better classifier for the next work.For getting good target swatches we do some necessary preprocessing before target recognition, such as resuming the rotating blurred image by using the local knowledge leading combined with a least cost function and adjusting the image because of the alterative size and transfer and illumination of targets.Secondly, the issue of Gabor Jets' parameters is researched. Two methods of chosing Gabor Jets' parameters are analyzed, and the best optimum Jets which extract the lest target feature information redundancy are obtained by optimizing parameters with the scatter matrix theory. The best parameters of Gabor Jets for extracting edge feature are sured.Thirdly, a new method using template matching thought is put forward to resolve dimension disaster produced by extracting feature using traditional Gabor Jets. It divides image into several layers and blocks before extracting Gabor feature. For enhancing the adaptability for the alter of target's lighteness and contrast farther, the extracted Gabor feature is improved. The experiment by Matlab programming is done on both standard Yale human face database and the self-making target database. And the result shows that the means which this paper put forward are more applicable and robust.Lastly, the moment technique is researched and improved for getting the better steady moment invariants. A dynamic link architecture and the Gabor Jets having no relationship with orientation are used to resolve the problem that it is difficult to match successfully when the target's gesture changes. The steady invariable feature is obtained with the feature fusing technique, that accounts for the problem that the target could not be recognised exactly when its gesture alters. The simulation experiment done on the self-making target database verify that the algorithm is applicable and universal.